All under github.com/virdx/. Read a repo’s AGENTS.md / CLAUDE.md before changing it.
| Repo | Owns | Go here when |
|---|---|---|
mono (packages/vxrun) |
The vxrun/kuberun CLI and vxrun.api: build, run, batch, workflow, status, describe, cancel, capacity, review, data inspect, storage. Released as a conda package on the internal Artifactory channel (vxrun 0.3.0 and later carry vxrun metrics and the Grafana path). |
A flag is missing, output is confusing, submission fails client-side. |
mono (apps/vx) |
The vx CLI and its MCP server (vx mcp locally, vx mcp serve centrally): tools, these guides, feedback. |
A tool here is wrong or a guide is stale (apps/vx/scripts/sync-docs.sh re-vendors submitting-jobs, the vxs3 guides, the SDK guide and the API bundle contract; cluster summaries remain hand-maintained). |
mono (packages/vxs3, packages/vxdata-sdk) |
Bundle format, storage analysis (vxs3 analyze), verified transfers and caching; authenticated vxdata transfers and Bundle publication (from vxdata.sdk import Client). |
Reading or publishing datasets from a job; inspect_data results. Read read_doc(“vx/cluster/vxs3”), read_doc(“vx/cluster/vxs3-storage”) and read_doc(“vx/cluster/vxdata-sdk”). |
mono (apps/vxdata-api, packages/vxdata-schemas) |
Catalog resource schemas, versioning, authorization and presigned storage access. | Bundle permissions, schema/migration requirements or API storage contracts; see read_doc(“vx/cluster/vxdata-api”). |
infra_k8s |
GitOps for the cluster: ArgoCD apps, namespaces, RBAC, Kueue queues and flavors, priority classes, Argo WorkflowTemplates (run-image, run-batch, builds), Kyverno policies (current/kyverno-policies/, current/argo-workflows/policies/), operator scripts in util/. SUBMITTING_JOBS.md is canonical for how jobs are submitted. |
Scheduling policy, quotas, templates, admission errors. |
infra_iam |
Source of truth for group membership, scheduling teams and quotas; renders the Kueue manifests and the team-membership ConfigMap in k8s/scheduling/. Files worth reading: data/scheduling/defaults.yaml (personal quota: 36 CPU / 264Gi / 2 GPUs, 10 concurrent Workflows), data/scheduling/teams.yaml (which teams have queues and their PRO/CPU/memory quota), data/groups/teams/<team>.yaml (members; a user may appear in several), data/users/<user>.yaml (per-user overrides). |
A user has no queue, needs more quota, asks which team they are in or what a team’s quota is, or a team should exist. |
infra_onboarding |
Onboarding automation: VPN, CA certificates, kube CLI login, Coder workspace setup. | A user cannot reach the cluster or Coder at all. |
infra_coder |
Coder workspace images. The dev image installs vxrun from the Artifactory channel (kuberun is a symlink to vxrun) and ships vx-update to pull the latest release without a rebuild. |
A tool is missing from the workspace, or the installed tools are behind the channel. |
kuberun |
Standalone predecessor of vxrun, superseded by mono/packages/vxrun on 2026-10-01. Not archived yet, but nothing installs from it any more. |
Only to compare behaviour during the transition. |
The team facts vendored into this server (names, member counts, quotas) are a snapshot from release time. For current answers clone infra_iam and read the files above, for example:
gh repo clone virdx/infra_iam /tmp/infra_iam -- --depth 1
grep -l "<username>" /tmp/infra_iam/data/groups/teams/*.yaml # which teams
cat /tmp/infra_iam/data/scheduling/teams.yaml # team quotasMembership changes are PRs to infra_iam (groups: [<team>] in the user’s
file); ArgoCD syncs the rendered manifests after merge.
Node-level problems (GPUs missing, NFS, DNS, VPN) are owned by the platform
team: report them with send_feedback (kind gap) rather than pointing users
at infrastructure repositories.
Monitoring lives in two places: the Prometheus rules, kube-state-metrics and
DCGM configuration in infra_k8s/current/monitoring/ and current/argo-cd/,
and Grafana itself (JWT auth, datasources, dashboards) in infra_ansible
(roles/grafana). See read_doc(“vx/cluster/metrics”).